Embeddings

Take the GloVe word vectors that Stanford released in 2014, trained on six billion tokens of Wikipedia and news text, and compute the most famous expression in the field: king − man + woman. Then ask which word vector lies closest to the result. In the 50-dimensional version the answer is not queen. It is king, with a cosine similarity of 0.885; queen comes second at 0.852. The celebrated result appears only if you remove the three input words from the candidates, which the standard analogy benchmarks quietly do.